Defense Notices

Depth inference is a fundamental problem of computer vision with a broad range of potential applications. Monocular depth inference techniques, particularly shape from shading dates back to as early as the 40's when it was first used to study the shape of the lunar surface. Since then there has been ample research to develop depth inference algorithms using monocular cues. Most of these are based on physical models of image formation and rely on a number of simplifying assumptions that do not hold for real world and natural imagery. Very few make use of the rich statistical information contained in real world images and their 3D information. There have been a few notable exceptions though. The study of statistics of natural scenes has been concentrated on outdoor natural scenes which are cluttered. Statistics of scenes of single objects has been less studied, but is an essential part of daily human interaction with the environment. This thesis focuses on studying the statistical properties of single objects and their 3D imagery, uncovering some interesting trends, which can benefit shape inference techniques. I acquired two databases: Single Object Range and HDR (SORH) and the Eton Myers Database of single objects, including laser-acquired depth, binocular stereo, photometric stereo and High Dynamic Range (HDR) photography. The fractal structure of natural images was previously well known, and thought to be a universal property. However, my research showed that the fractal structure of single objects and surfaces is governed by a wholly different set of rules. Classical computer vision problems of binocular and multi-view stereo, photometric stereo, shape from shading, structure from motion, and others, all rely on accurate and complete models of which 3D shapes and textures are plausible in nature, to avoid producing unlikely outputs. Bayesian approaches are common for these problems, and hopefully the findings on the statistics of the shape of single objects from this work and others will both inform new and more accurate Bayesian priors on shape, and also enable more efficient probabilistic inference procedures.

The work presented in this thesis explores the use of geographic data and geostatistical methods to estimate path loss for cognitive radio networks. Path loss models typically employed in this scenario use a general terrain type (i.e., urban, suburban, or rural) and possibly a digital elevation model to predict excess path loss over the free space model. Additional descriptive knowledge of the local environment can be used to make more accurate path loss predictions. This research focuses on the use of visible imagery, digital elevation models, and terrain classification systems for predicting localized propagation characteristics. A low-cost data collection platform was created and used to generate a sufficiently large spectrum measurement set for machine learning. A series of path loss models were fitted to the data using linear and nonlinear methods. These models were then used to create a radio environment map depicting estimated signal strength. All of the models created have good cross-validated prediction results when compared to existing path loss models, although some of the more flexible models had a tendency to overfit the data. A number of geostatistical models were fitted on the data as well. 
These models have the advantage of not requiring the transmitter location in order to create a model. The geostatistical models performed very well when given a sufficient density of observations but were not able to generalize as well as some of the regression models. An analysis of the geographical data sets indicated that each had a significant measurable effect on path loss estimation, with the medium resolution imagery and elevation data providing the greatest increase in accuracy. Finally, these models were compared to number of existing path loss models, demonstrating a gain in usable spectrum for cognitive radio network use.

Tactical networks, networks designed to facilitate command and control capabilities for militaries, have key attributes that differ from the commercial Internet. Characterizing, modeling, and ex- ploiting our understanding of these differences is the focus of this research. 
The differences between tactical and commercial networks can be found primarily in the areas of access bandwidth, access diversity, access latency, core latency, subnet distribution, and network infrastructure. In this work we characterize and model these differences. These key attributes affect research into issues such as peer-to-peer protocols, service discovery, and server selection among others, as well as the deployment of services and systems in tactical networks. Researchers traditionally struggle with measuring, analyzing, or testing new ideas on tactical networks due to a lack of direct access, and thus this characterization is crucial to evolving this research field. 
In this work we develop a topology generator that creates realistic tactical networks that can be visualized, analyzed, and emulated. 
Topological features including geographically constrained line of sight networks, high density low bandwidth satellite networks, and the latest high bandwidth on- the-move networks are captured. All of these topological features can be mixed to create realistic networks for many different tactical scenarios. A web based visualization tool is developed, as well as the ability to export topologies to the Mininet network virtualization environment. 
Finally, state-of-the-art server selection algorithms are reviewed and found to perform poorly for tactical networks. We develop a collaborative algorithm tailored to the attributes of tactical networks, and utilize our generated networks to assess the algorithm, finding a reduction in utilized bandwidth and a significant reduction in client to server latency as key improvements.

Multitask Learning is a learning framework which explores the concept of sharing training information among multiple related tasks to improve the generalization error of each task. The benefits of multitask learning have been shown both empirically and theoretically. There are a number of fields that benefit from multitask learning, including toxicology. However, the current multitask learning algorithms make a very important key assumption that all the tasks are related to each other in a similar fashion in multitask learning. The users often do not have the knowledge of which tasks are related and train all tasks together. This results in sharing of training information even among the unrelated tasks. Training unrelated tasks together can cause a negative transfer and deteriorate the performance of multitask learning. For example, consider the case of predicting in vivo toxicity of chemicals at various endpoints from the chemical structure. Toxicity at all the endpoints are not related. Since, biological networks are highly complex, it is also not possible to predetermine which endpoints are related. Thus, training all the endpoints together may cause a negative effect on the overall performance. This proposal aims at developing algorithms which make use of task relationship models to further improve the generalization error and prevent transfer of information among the unrelated tasks. The algorithms proposed here either learn the task relationships or utilize the known task relationships in the learning framework. Further, these algorithms will be utilized to predict toxicity of chemicals at various endpoints using the chemical structures and the results of multiple in vitro assays performed on these chemicals.

A rule set is a popular symbolic representation of knowledge derived from 
data. A rule induction is an important technique of data mining or machine 
learning. Many rule induction algorithms are widely used, such as LEM1, LEM2 and MLEM2. Some of these algorithms perform better on special data, e. g., on inconsistent data set or data sets with missing attribute values. This work discusses basic ideas of the MLEM2 algorithm, especially, how it handles data sets with numeric attribute values. Additionally, a comparison of the performance of different discretization options of the MLEM2 algorithm is also included.

The large data stores of Personally Identifiable Information (PII) in todays connected systems, coupled with the increased potential damages of Identity Theft bring the need for architectures that provide secure collection, storage, and transmission of this data. The need has not yet been standardized in the industry in a way similar to the Payment Card Industry (PCI) has done so. At the same time, however, municipalities, states, and even countries are instituting legislature that requires business entities that store PII data to maintain adequate security of the data. The need has become clear for a set of processes, procedures, and systems that provide a framework for securely storing PII data. This project defines the lower level datastore system and associated services for that PII data. It also outlines a network architecture prototype for providing segmented security zones used to provide more layers of security in a connected system.

We all consult online reviews before obtaining a product or service. However, not all the reviews can be trusted. For example, in 2013, "Operation Clean Turf” a yearlong sting operation in New York State, caught 19different companies that were writing fake reviews in online forums like Yelp for businesses that paid them. For my project, I've developed an application called AskMyNetwork. AskMyNetwork interfaces with Facebook to obtain feedback or input from a user's Facebook friends.The rationale for my project is that the feedback or inputs are from "friends" (personal friends, family members,or colleagues in a user's Facebook friends' list) and can be trusted. 

AskMyNetwork has four major components namely, Login,Search My Network, Ask My Network and Notifications. Using the Login component, the user can login to the application with Facebook credentials. Using Search My Network component, the user can define search criteria (e.g.,search for restaurant in Kansas City) and search his or her Facebook data for relevant results. Using Ask My Network component, the user can ask a group of friends question about a product or service they would like an opinion on. The group of friends can either be chosen by name or by the current location of the friends. Using the Notifications component, the user can view the responses given to questions asked from AskMyNetwork. 

I validated AskMyNetwork via a number of inquiries on topics such as restaurants, places to visit in a city and arts. The results of the validation were satisfactory.

Power line communication (PLC) has received steady interest over recent decades because of its economic use of existing power lines, and is one of the communication technologies envisaged for Smart Grid infrastructure. However, power lines are not designed for data communication, and this brings unique challenges for data communication over power lines. In particular for broadband (BB) PLC, the channel exhibits linear periodically time varying (LPTV) behavior synchronous to the AC mains cycle due to time varying impedances, impulsive noise due to switching events in the power line network is present in addition to background noise. In this work, we focus on two major aspects of an orthogonal frequency division multiplexing (OFDM) system for BB PLC LPTV channels; bit and power allocation, and channel estimation (CE). 

For the problem of optimal bit and power allocation, we present that the application of a power constraint that is averaged over many microslots can be exploited for further performance improvements through bit loading. Due to the matroid structure of the optimization problem, greedy-type algorithms are proven to be optimal for the new LPTV-aware bit and power loading. Next, two mechanisms are utilized to reduce the complexity of the optimal LPTV-aware bit loading and peak microslot power levels: employing representative values from microslot transfer functions, and power clipping. 

Next, we introduce a robust CE scheme with low overhead that addresses the drawbacks of block-type pilot arrangement and decision directed CE schemes such as large estimation overhead, and difficulty in channel tracking in the case of sudden changes in the channel, respectively. A transform domain (TD) analysis approach is developed to determine the cause of changes in the channel estimates. The result of TD analysis is then exploited in the proposed scheme to mitigate the effects of LPTV channel and impulsive noise. 

Our results indicate that the proposed reduced complexity LPTV-aware bit loading with power clipping algorithm performs close to the optimal scheme, and the proposed CE scheme based on TD analysis has low estimation overhead and is robust to changes in the channel and noise, making them good alternatives for BB PLC LPTV channels.

Many of the innovative algorithms that permeate the field of array processing are based on a very simple signal model of an array. This simple, although powerful, model is at times a pale reflection of the complexities inherent in the physical world, and this model mismatch opens the door to the performance degradation of any solution for which the model underpins. This dissertation seeks to explore the impact of model mismatch upon common array processing algorithms. Model mismatch is examined in two ways: First, by developing a blind array calibration routine that estimates model mismatch and incorporates that knowledge into the RISR direction of arrival estimation algorithm. Second, by examining model mismatch between a transmitting and receiving antenna array, and assessing the impact of this mismatch on prolific direction of arrival estimation algorithms. In both of these studies it is shown that engineers have traded algorithm performance of model simplicity, and that if we are willing to deal with the added complexity we can recapture that lost performance.

A method of transmitting information clandestinely over a variety of network protocols is designed and discussed. A demonstrative implementation is created that utilizes the ubiquitous Hypertext Transfer Protocol (HTTP) and the world wide web. Key contributions include the use of access ordering to convey information, and the modulation of transaction level timing to emulate user behavior.